Automatic Fixation of Decompilation Quirks Using Pre-trained Language Model

Decompiler is a system for recovering the original code from bytecode. A critical challenge in decompilers is that the decompiled code contains differences from the original code. These differences not only reduce the readability of the source code but may also change the program's behavior. In this study, we propose a deep learning-based quirk fixation method that adopts grammatical error correction. One advantage of the proposed method is that it can be applied to any decompiler and programming language. Our experimental results show that the proposed method removes 55% of identifier quirks and 91% of structural quirks. In some cases, however, the proposed method injected a small amount of new quirks.